Mitigating crop modeling uncertainties through machine learning in drylands

Dry farming systems in drylands are sensitive to climate extremes, resulting in large yield variability and posing challenges to food security. Reliable crop modeling approaches are therefore essential, particularly under data-scarce conditions. This study evaluated the performance of the CSM-CERES-Wheat model across selected drylands in Iran using five gridded meteorological datasets. CFS and ERA5-Land provided daily precipitation, temperature and solar radiation, while CHIRPS, IMERG and PERSIANN-CDR supplied only precipitation. A novel clustering-unbiasing-ensembling framework was applied using four machine learning (ML) algorithms. Light Gradient Boosting Machine and Random Forest outperformed other MLs, particularly for precipitation and minimum temperature. These top-performing MLs increased the average Nash-Sutcliffe Efficiency from 0.11-0.17 to 0.47 for precipitation and from 0.56-0.88 to 0.89-0.96 for temperature and solar radiation. Correcting variables individually had limited impact, whereas the TOTAL scenario-combining all corrected variables-improved yield and water stress simulations in ~ 60% of cases. Bootstrapping confirmed that the TOTAL scenario results were reliable, thereby supporting its transferability. Furthermore, ML-based TOTAL corrections outperformed classical unbiasing-ensembling algorithm. These findings underscore the value of advanced ML-based unbiasing-ensembling for enhancing climate data reliability in crop modeling, offering practical guidance to support susceptible dryland agricultural systems under extremes.